Here’s an example of how a single word can make or break your AI patent claim.
During my epi webinar “Patenting AI-Related Inventions” yesterday, I discussed a case study involving L’Oréal’s famous patent application EP 4 000 000. Claim 1 was a runtime method directed at using a machine learning model to estimate the skin color of a face in an image.
To avoid a lack of support rejection, the EPO examiner forced L’Oréal to specify exactly how the model was trained. This prompted a brilliant question from one of the webinar participants:
“Doesn’t adding the training to a runtime method create a divided infringement issue?”
And this is where claim drafting becomes an art form.
If L’Oréal had added the training process as active method steps, it absolutely would have created a divided infringement hurdle. Why? Because the entity training the model is rarely the same entity executing the trained model.
But they (or their representative; hat tip to the colleagues at CASALONGA) were clever. They added the training details in a “wherein” clause. The resulting claim listed the active runtime steps (receiving an image, applying the model, outputting the result) and simply concluded with: “wherein the machine learning model is trained by…”
By doing this, the claim only requires that the model is in a trained state (by having undergone the recited training process). It does not require the execution of those training steps.
To a data scientist, semantics like this probably sound bizarre. But from a legal standpoint, it is mission-critical. One formulation makes you lose a multi-million-dollar infringement case; the other makes you succeed.
What is the smallest word change you’ve seen that made the biggest difference in a claim’s scope?
Hope this helps,
Bastian